Which AI level does this job need?
The useful move is choosing the lowest level that removes the bottleneck.
Once you see the AI ladder — for context, go check yesterday’s post — the next step is to identify what to use each level for.
Chat is best for quick, disposable thinking: research starting points, quick questions, definitions, news context, learning a new topic, rewriting a paragraph, pressure-testing an idea.
Although, chat still needs a good prompt.
Personalized assistants are best for repeated work with a familiar format.
Proposal drafts. Meal plans. Client emails. Content repurposing. Internal SOPs. Even Substack thumbnails (👀). Anything where the task repeats and the AI benefits from knowing your tone, examples, rules, and preferred structure.
Embedded assistants are best when the context already lives on your filesystem.
Code is the obvious and common example, but the same idea applies to CSV exports, meeting notes, project documents or files generated by other systems.
The advantage is access. Instead of uploading files manually, the AI can work inside the environment.
Also, embedded tools increasingly integrate with the other apps from your stack, hence multiplying leverage.
Agents are best for goals that need multiple steps.
Automated meeting notes collection and processing. Scheduled reporting. External data gathering. Research that runs on a cadence. Event-based workflows.
Anything where the AI needs to watch, collect, process, and push the next step forward.
Of course, levels overlap.
A personalized assistant can use files. Embedded assistants have agentic features like scheduled tasks. Agents include memory and workspace access by default.
And that’s fine, the point is not classify every tool perfectly, but to stop using the same tool for everything.
If you need help increasing your technical clarity to create more leverage, let’s talk.